advanced-evaluation
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill architecture is designed to ingest and process untrusted LLM outputs, which are vulnerable to embedded instructions that could attempt to influence the judge's scoring logic.
- Ingestion points: Untrusted content is interpolated into evaluator prompts using placeholders such as
{response},{response_a}, and{response_b}(documented inSKILL.mdunder 'Prompt Structure' sections). - Boundary markers: The skill utilizes clear Markdown header delimiters (e.g.,
## Response to Evaluate) to separate instructions from untrusted data, reducing the likelihood of successful injection. - Capability inventory: The provided scripts and patterns focus on text analysis and scoring via the LLM judge. No evidence of high-privilege operations (file system writes, arbitrary command execution) being triggered by these inputs was found in the provided code.
- Sanitization: While the skill promotes structured JSON schemas for outputs to prevent format-breaking attacks, it does not explicitly detail input sanitization or escaping for the ingested content.
Audit Metadata